Research map / Vision-language and multimodal / Human and embodied modeling
LLaMo: Scaling Pretrained Language Models for Unified Motion Understanding and Generation with Continuous Autoregressive Tokens
Zekun Li, Sizhe An, Chengcheng Tang, Chuan Guo, Ivan Shugurov, Linguang Zhang, Amy Zhao, Srinath Sridhar, Lingling Tao, Abhay Mittal
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use modality-specific MoT with frozen text modules instead of full LLM fine-tuning (3B) (ablation) | LLaMo · MMLU | accuracy: 24.9 -> 63.4 | Improved |
| use MoT with frozen text modules instead of full fine-tuning (1B) (ablation) | LLaMo · MMLU | accuracy: 26.6 -> 49.3 | Improved |
| use MoT instead of full fine-tuning (3B) (ablation) | LLaMo · Own large-scale motion-text test split (~30K) | FID: 46.174 -> 19.893 | Improved |
| use MoT instead of full fine-tuning (1B) (ablation) | LLaMo · Own large-scale motion-text test split (~30K) | FID: 63.215 -> 27.361 | Improved |
| use MoT instead of full fine-tuning (3B) (ablation) | LLaMo · Own large-scale motion-text test split (~30K) | text-to-motion R@3: 0.8307 -> 0.9594 | Improved |
| use MoT instead of full fine-tuning (3B) (ablation) | LLaMo · Own large-scale motion-text test split (~30K) | motion-to-text R@3: 0.9412 -> 0.9422 | No clear effect |
| scale LLaMo from 1B to 3B (ablation) | LLaMo · HumanML3D text-to-motion | FID: 53.942 -> 22.491 | Improved |
| scale LLaMo from 1B to 3B (ablation) | LLaMo · HumanML3D motion-to-text | CIDEr: 104.7 -> 100.8 | No clear effect |
| use continuous causal VAE instead of FSQ discrete tokenizer (ablation) | motion codec · Motion reconstruction | MPJPE: 41.9 -> 10.1 | Improved |
| increase VAE latent dimension from 32 to 64 (ablation) | causal VAE · Motion reconstruction | MPJPE: 10.1 -> 3.86 | Improved |
| LLaMo vs specialist models on HumanML3D FID | LLaMo-3B vs MotionStreamer · HumanML3D text-to-motion | FID: 11.790 -> 22.491 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- MotionGPT3 — MotionGPT: Human Motion as a Foreign Language (NeurIPS 2023) · better on 1 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion (CVPR 2025 · similarity 0.90)
- MG-MotionLLM: A Unified Framework for Motion Comprehension and Generation across Multiple Granularities (CVPR 2025 · similarity 0.90)
- Human Motion Instruction Tuning (CVPR 2025 · similarity 0.89)
- LaMP: Language-Motion Pretraining for Motion Generation, Retrieval, and Captioning (ICLR 2025 · similarity 0.89)
- HSI-GPT: A General-Purpose Large Scene-Motion-Language Model for Human Scene Interaction (CVPR 2025 · similarity 0.89)
- M$^3$GPT: An Advanced Multimodal, Multitask Framework for Motion Comprehension and Generation (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
